Gast ::
| Hauptseite > Publikationsdatenbank > Autonomous Image Analysis to Accelerate the Discovery and Integration of Energy Materials |
| Hauptseite > Publikationsdatenbank > Autonomous Image Analysis to Accelerate the Discovery and Integration of Energy Materials |
| Book/Dissertation / PhD Thesis | FZJ-2026-02425 |
2026
Forschungszentrum Jülich GmbH Zentralbibliothek, Verlag
Jülich
ISBN: 978-3-95806-911-4
This record in other databases:
Please use a persistent id in citations: urn:nbn:de:0001-2606301402027.961192643455 doi:10.34734/FZJ-2026-02425
Abstract: Accelerating the discovery and development of advanced energy materials is critical to transitioning to a sustainable clean energy future. Data-driven methods, especially those using artificial intelligence (AI) and deep learning (DL), offer unprecedented exploration opportunities to significantly increase the speed at which new energy materials can be explored and optimized. Central to materials innovation and understanding of materialsrelated phenomena are advanced characterization techniques, in which imaging plays a vital role. Interpreting these datasets manually remains time-consuming and humanbiased. To fully realize the potential of data-driven techniques, it is essential to accelerate the analysis via artificial intelligence and a standardized and efficient system for managing imaging data. This work demonstrates the research acceleration capabilities of using DL to automate image data analysis. For this purpose, highly relevant use case scenarios were investigated with a focus on energy-related electrochemical systems. First, in the context of polymer electrolyte water electrolyzers (PEWEs), a 2D DL-based framework was developed that is tailored for high-throughput analysis of optical video recordings of the oxygen bubble evolution of a transparent cell. After binary bubble segmentation, the automated software provides quantitative insights into bubble dynamics, including time-resolved bubble coverage evolution, size distributions, bubble density maps, and morphological analysis. These results significantly improve the experimental understanding of oxygen gas dynamics in PEWE cells under different conditions, providing faster data extractions previously inaccessible due to inefficient traditional methods. Building on these 2D bubble-analysis results, the challenges of 3D bubble analysis in vanadium redox flow batteries (VRFBs) are addressed through automated DL-based analysis of 3D synchrotron X-ray tomograms. A multi-class semantic segmentation approach was developed using a comprehensive dataset of 2294 annotated images from three different battery configurations to distinguish between bubble, electrolyte, membrane, and gasket classes with excellent performance. The developed tool addressed the challenging task of accelerating high-throughput volume analysis and streamlining bubble quantification processes. The software enables automatic feature extraction of bubble volumes, shapes, and membrane blockage. An interactive 3D visualization tool was developed to improve the visual inspection of the analyzed volumes and their properties on the fly. Having demonstrated 3D bubble segmentation, 3D architectures were employed to tackle complex porous-material segmentation in PEM technologies. The complexity of accurately characterizing gas diffusion layers (GDLs), microporous layers (MPLs), and catalyst layers (CLs) typically renders traditional manual and algorithmic methods inefficient. Therefore, a DL framework for segmenting micro-CT and FIB-SEM characterized porous volumes was developed and validated with physical porosity measurements. Using state-of-the-art 3D neural networks, the software successfully achieved robust binary pore/material segmentation and multi-class pore/GDL/MPL segmentation. This segmentation tool provides case-specific quantification capabilities for pore size distributions, porosity, MPL crack analysis, and MPL intrusion assessment. In addition, the 3D segmentation volumes can be visualizedin situ employing the developed visualization capabilities. Based on the experience gained in these contributions, developing an infrastructure designed for the standardized storage and retrieval of imaging and characterization data from energy materials experiments was crucial. Using an adapted version of the elementary multiperspective materials ontology (EMMO), a standardized metadata schema has been developed to capture the essence of different imaging modalities related to energy materials. This was achieved by combining the expertise of seven experts in various fields. This approach ensures that imaging datasets are universally stored in a semantically consistent format, as FAIR guidelines require. In addition, the work implements the metadata schema into a native graph database integrated with an ontology alignment system based on large language models (LLMs). As a result, the developed methods bridge the gap between FAIR principles and a practical implementation for experimentalists. Together, these milestones form an end-to-end DL pipeline for automated, high-throughput analysis of electrochemical systems and the subsequent data storage. The open-access tools developed through research included in this thesis represent a significant advancement in data-driven image analysis and standardized data management. Furthermore, the expertise gained from this work has allowed the development of standardized metadata protocols to facilitate data interoperability and discoverability, accelerating the further training of DL models. Finally, combining both approaches opens a path for conceptualizing a unified platform that benefits from the developed metadata approach for data storage to create an accelerated materials research environment.
|
The record appears in these collections: |
PDF
Fulltext by OpenAccess repository